Research On Fabric Image Retrieval Based On Fine-grained Features

Fabric image retrieval technology not only has potential application value in visual tasks such as fabric recognition and recommendation system,but also has important application value in many fields such as e-commerce,fabric inventory management,textile product design,etc.,so it has become a research hotspot in recent years.However,due to the diversity of textile fabric appearance and the fineness of fabric texture,it is difficult to extract the fine-grained features of fabric images during fabric retrieval.In order to overcome the shortcomings of the current fabric image feature extraction technology,such as low feature expression ability,only focusing on the overall feature,and long network training time.Relying on the fabric image database collected by the research group,this paper uses fine-grained features to realize the precise retrieval of fabric images.The specific work is as follows:(1)A fabric image feature extraction model based on fine-grained features is proposed.Aiming at the problem that the features extracted by most convolutional networks are the macroscopic features of fabric images and lack of highlighting the fine-grained features of fabric images,a feature extraction model based on fine-grained features is proposed.By introducing the coordinate attention module to strengthen the network’s attention to the fine texture,weaving structure and other local details of the fabric image,the image retrieval accuracy is further improved.The experimental results show that the precision and recall of the network with the coordinate attention module have reached 91.0% and 89.6%,which are 15.7%and 16.0% higher than the original MobileNetV3 network.(2)A fabric image retrieval speed-up model based on the improved MobileNetV3 network is proposed.Although the MobileNetV3 network optimizes the computationally intensive layer structure in the MobileNetV2 network,it reduces the burden of network training by reducing the number of convolution kernels and reducing the calculation of the three convolutional layers.But for fabric image retrieval,the amount of parameters is still too large.Inspired by the Efficient Net network,the scaling factor method is used to scale the network structure of MobileNetV3 in terms of width and height to reduce the number of model parameters,reduce network training time,and improve retrieval speed.Experiments show that,compared with the benchmark MobileNetV3 model,the structure-optimized network has an output parameter quantity of 3.74 M and FLOPs of 175.34 M while keeping the retrieval accuracy approximately unchanged,reducing the network’s calculation load by 1.74 M.At the same time,the average retrieval time is 1.8s,and the retrieval speed is increased by 65%.(3)A fabric image retrieval model based on feature fusion—FA-MobileNet is proposed.Aiming at the problem that the fine-grained features extracted by the convolutional neural network cannot contain the color information of the fabric image,this paper uses the color histogram to extract the color features of the fabric image and fuses them with the fine-grained features to improve the retrieval performance.The experimental results show that the fusion strategy can effectively improve the performance of fabric image retrieval,and the precision rate and Top-5 accuracy rate of the fabric image retrieval model are 92.63% and 75.65%.It is0.81% and 4.27% higher than the improved MobileNetV3 network without feature fusion.(4)Finally,a fabric image retrieval system is constructed using a fabric image retrieval model based on feature fusion.The system can quickly and effectively realize the image retrieval of textile fabrics focusing on color combined with fine-grained features,which significantly improves the production efficiency and competitiveness of textile enterprises,meets the needs of practical applications,and has broad application prospects.